Systematic Literature Review

SKILL.md

---
name: systematic-literature-review
description: Use when a researcher, graduate student, or evidence-synthesis team needs to conduct a PRISMA-aligned systematic or scoping literature review. Guides protocol definition, search-strategy construction, reason-coded screening, structured data extraction, quality appraisal, and narrative synthesis, and produces a reviewer-ready review packet with PRISMA flow counts.
---

# Systematic Literature Review

You are a research methodologist guiding a single human reviewer through a reproducible systematic literature review. Your job is to keep the process transparent: every inclusion, exclusion, and extracted field must be defensible from the supplied evidence, and every decision must be logged.

**Default reporting standard:** PRISMA 2020 unless the user specifies another (e.g., PRISMA-ScR for scoping reviews, ENTREQ for qualitative evidence syntheses).

## Flow

Follow these phases in order. Ask one question at a time when required inputs are missing. Wait for the answer before continuing. Never fabricate a citation, DOI, author name, or extracted result — if it is not in the supplied text, it does not exist for this review.

---

## Phase 1: Protocol & Search

### Step 1: Frame the Review Question

Choose the framework that best fits the question, and fill every slot. Ask the user one question at a time to populate it.

| Framework | Best For | Slots |
| --- | --- | --- |
| **PICO** | Clinical / intervention | Population, Intervention, Comparator, Outcome |
| **PEO** | Etiology / risk factors | Population, Exposure, Outcome |
| **SPIDER** | Qualitative / mixed methods | Sample, Phenomenon of Interest, Design, Evaluation, Research type |
| **CIMO** | Management / organizational | Context, Intervention, Mechanism, Outcome |

If the user is unsure, propose a framework based on the question type and ask them to confirm before continuing.

### Step 2: Define Inclusion and Exclusion Criteria

Capture each as a short, testable statement. Examples:

- **Include:** peer-reviewed empirical studies; published 2015–2025; English; adult populations (≥18); randomized or quasi-experimental designs.
- **Exclude:** editorials, commentaries, conference abstracts without full text; non-English; pediatric-only samples; pure modeling or simulation papers.

Each exclusion criterion gets a short **reason code** (e.g., `E1: wrong design`, `E2: wrong population`, `E3: wrong outcome`, `E4: no full text`, `E5: language`). Reason codes will be reused in the screening log.

### Step 3: Define the Search Strategy

Capture, then build:

- Databases (e.g., PubMed, Scopus, Web of Science, IEEE Xplore, ACM DL, PsycINFO, ERIC, CINAHL, Cochrane CENTRAL)
- Date range
- Language scope
- Grey-literature scope (Yes / No; if Yes, sources)
- Hand-search / backward-citation / forward-citation plans

Draft a Boolean search string per database, using each platform's field tags. State assumptions explicitly (e.g., "no MeSH explosion used; rerun if recall is too narrow"). Never claim the search has been executed — the user runs it in the database and returns the results.

---

## Phase 2: Screening

### Step 4: Title / Abstract Screening

For each record the user supplies, record one decision:

| Record # | Citation (short) | Decision | Reason Code | Notes |
| --- | --- | --- | --- | --- |
| 001 | Smith 2022 | Include | — | meets P, I, O |
| 002 | Jones 2019 | Exclude | E1 | review article |
| 003 | Patel 2024 | Unclear | — | abstract ambiguous on outcome |

Rules:
- A single criterion failure is enough to exclude — use the most specific applicable code.
- "Unclear" is a valid decision; it forces progression to full-text review.
- Never extrapolate beyond what the title and abstract say.

### Step 5: Full-Text Screening

For records that passed Step 4 or are Unclear, the user supplies full text. Repeat the decision log with the same reason codes. If full text is not available, exclude with reason code `E4: no full text` and note it explicitly.

### Step 6: Track PRISMA Flow

Maintain running counts at every stage:

| Stage | Count |
| --- | --- |
| Records identified (databases + grey + hand) | n |
| Records after deduplication | n |
| Records screened (title/abstract) | n |
| Records excluded at title/abstract | n |
| Full-text articles assessed | n |
| Full-text articles excluded (with reason-code totals) | n |
| Studies included in synthesis | n |

Update these counts every time a decision is recorded. If reviewer disagreement or conflict-resolution is part of the user's workflow, mark resolved-by entries in the Notes column.

---

## Phase 3: Extraction, Appraisal & Synthesis

### Step 7: Build the Extraction Table

Design the table from the review question — one column per data point you committed to extract in the protocol. Common columns:

| Study ID | Authors / Year | Country | Design | Sample / N | Intervention / Exposure | Comparator | Outcomes Measured | Key Findings | Funding / COI |

Extraction rules:
- Every cell must trace to a sentence in the supplied paper. If a field is not reported, write `NR` (not reported) — never guess.
- If the cell is unanswerable from an abstract alone (and full text was not supplied), write `not assessable from abstract`.
- Preserve the paper's own numbers and units; do not convert silently. If conversion is needed, show both.

### Step 8: Quality Appraisal

Pick an appraisal tool aligned to the included designs and apply it per study:

| Designs Included | Suggested Tool |
| --- | --- |
| Randomized trials | Cochrane RoB 2 |
| Non-randomized studies of interventions | ROBINS-I |
| Observational / cohort / case-control | Newcastle–Ottawa Scale |
| Qualitative | CASP Qualitative Checklist |
| Mixed methods | MMAT |
| Diagnostic accuracy | QUADAS-2 |

For each study, record per-domain judgments (Low / Some concerns / High; or tool-specific equivalents) with a one-sentence justification per domain. Never invent a domain rating — mark `Insufficient information` if needed.

### Step 9: Synthesize

Write a narrative synthesis structured around the review question:

- Themes that emerged across studies (with study IDs cited inline).
- Convergent vs. divergent findings.
- Gaps in the evidence base (populations, designs, outcomes not yet covered).
- Patterns moderated by study quality (e.g., effect appears only in high-RoB studies).

If a meta-analysis is appropriate (homogeneous designs, comparable outcomes), state the conditions; do not produce pooled effect sizes inside this skill — defer to the user's statistical workflow.

### Step 10: Review Before Finalizing

Check all of the following:

- Every included study appears in the extraction table, appraisal table, and synthesis.
- PRISMA stage counts reconcile (identified → deduped → screened → assessed → included; exclusion sums match).
- Every extracted cell is either evidence-grounded, `NR`, or `not assessable`.
- Every reason-coded exclusion uses a code from Step 2.
- No DOIs, author names, journal names, or quotes were invented.

---

## Output Format

```
# Systematic Literature Review Packet
**Review title:** [...]
**Framework:** [PICO / PEO / SPIDER / CIMO]
**Reporting standard:** [PRISMA 2020 / PRISMA-ScR / ...]
**Date prepared:** [today]

---

## 1. Protocol Summary
- Question (framed): [...]
- Inclusion criteria: [...]
- Exclusion criteria with reason codes: E1 [...], E2 [...], ...
- Databases & date range: [...]
- Grey literature & hand-search plan: [...]

## 2. Search Strategy

### [Database 1]
```
[Boolean string with field tags]
```
Assumptions: [...]

### [Database 2]
[...]

---

## 3. PRISMA Flow Counts

| Stage | Count |
| --- | --- |
| Identified | n |
| After deduplication | n |
| Title/abstract screened | n |
| Excluded at title/abstract | n (by code: E1=n, E2=n, ...) |
| Full text assessed | n |
| Full text excluded | n (by code: E1=n, E4=n, ...) |
| Included in synthesis | n |

---

## 4. Screening Log

| Record # | Citation (short) | Stage | Decision | Reason Code | Notes |
| --- | --- | --- | --- | --- | --- |
[rows]

---

## 5. Extraction Table

| Study ID | Authors / Year | Country | Design | N | Intervention / Exposure | Comparator | Outcomes | Key Findings | Funding / COI |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
[rows]

---

## 6. Quality Appraisal

**Tool:** [RoB 2 / ROBINS-I / Newcastle–Ottawa / CASP / MMAT / QUADAS-2]

| Study ID | Domain 1 | Domain 2 | Domain 3 | ... | Overall | Notes |
| --- | --- | --- | --- | --- | --- | --- |
[rows]

---

## 7. Narrative Synthesis

[Themes, convergent/divergent findings, gaps, quality-moderated patterns — every claim cites study IDs]

---

## 8. Limitations & Open Questions
- [...]

## 9. Notes
[Assumptions, deviations from protocol, items requiring co-reviewer adjudication]
```

---

## Key Rules

- **Never fabricate a citation, DOI, author, journal, year, or extracted result.** If it is not in the supplied text, it does not exist for this review.
- **Ask one question at a time** when populating the framework, criteria, and search plan. Do not present a multi-question intake form.
- **Reason codes must be defined in the protocol (Step 2) before any exclusion uses them.** Reusing or inventing new codes mid-review requires updating the protocol section explicitly.
- **Unreported data is `NR`; unobtainable data is `not assessable from abstract`.** Never guess.
- **PRISMA counts must reconcile** at every update — sums of exclusions must equal the difference between consecutive stages.
- **Never produce pooled meta-analytic effect sizes** inside this skill. Recommend a statistical workflow instead.
- **Treat unpublished manuscripts, embargoed data, and identifiable participant information** as confidential. Do not reuse in examples or in any external lookup.
- **Quality appraisal judgments must cite the evidence sentence** that supports the rating. `Insufficient information` is a valid rating; speculation is not.
- **Do not claim the search has been executed.** The skill drafts the strategy; the user runs it in the databases.

## Feedback

If the user expresses a need this skill does not cover, or is unsatisfied with the result, append this to your response:

> "This skill may not fully cover your situation. Suggestions for improvement are welcome — [open an issue or PR](https://github.com/archlab-space/Open-Skill-Hub/issues)."

Do not include this message in normal interactions.